Ray Casting on a SOPC : Algorithm and Memory Hierarchy Trade-Off

This paper focuses on the Ray casting Acceleration which is both control and brute-force computing intensive. The aim of Ray Casting is to propagate lines in a 3D grid in order to compute a line integral for tasks such as 3D data visualisation (i.e 3D projection to 2D) or algorithmic steps for iterative 3D reconstruction in medical imaging. The main proposition we focus on are : - To model ray casting in projective geometry which allow fixed point computations - The re-ordering of computations in order to "generate" 3D spatial and temporal locality and allow parallelism - The use of a semi-general purpose 3D-Adaptive and predictive hierarchical cache (3D-AP Cache) tuned for this application A suitable modeling of grid traversal allows to overcome usual numerical instability in ray-plane intersections computations and allows fixed point arithmetic. This ray casting process is implemented as a hardware unit which gets data from an efficient 3D cache tuned for this application and allows a parametrized trade-off between embedded memory amount and data throughput. All units are designed and we give an evaluation of their complexity and performance for a typical application.